Fine-Tuning Pretrained Deep Learning Models for Multi-Class Chest X-Ray–Based Pulmonary Disease Prediction: A Controlled Evaluation

Taking into account the common practice of benchmarking multiple pretrained models and selecting a single best-performing architecture, this study examines whether any model consistently outperforms others across different pulmonary disease categories. We employ a unified evaluation framework in which several state-of-the-art pretrained models, including ResNet50, MobileNet, DenseNet, EfficientNet, Vision Transformer (ViT), and MaxViT, are fine-tuned and evaluated on the same chest X-ray dataset. The results show that no single model achieves superior performance across all diseases and evaluation criteria. Instead, model effectiveness is disease-dependent and influenced by clinically relevant factors such as recall, false negative rate, and specificity. While transformer-based architectures perform well for certain conditions, convolutional models demonstrate advantages in others. These findings highlight the limitations of single-model selection strategies and support parallel multi-model evaluation for capturing diverse pathological patterns. Although this approach increases computational cost, it enables more clinically informed and robust model selection for pulmonary disease prediction.